handle special action post-conditions

Close environment

Collect a single trajectory. Returns: Dictionary containing the full trajectory data.

Strip HTML of tags and add separators to convert observation into simple mode

Convert padding structure and return sorted tensor with indices.

Convert raw data to training format expected by the system. Each item contains the task and expected answer.

Create simple wrapper functions for common tools.

decompress compressed mask array

Auto-discover all available tools in the tools directory.

Render and return HTML for done page

Conduct dynamic rollouts until a target batch size is met. Keeps sampling until the desired number of effective trajectories is coll

Compute episode-level advantage using mean-std normalization for GiGPO. (with only one scalar reward for each episode). Args: tok

Run per image evaluation on given images and store results (a list of dict) in self.evalImgs :return: None

Evaluate agent answers against ground truth using LLM judge. Returns: Dict containing: - individual_

Fetches the latest news articles from Nature's website. Parameters: num_articles (int): The number of articles to fetch.

Execute a search and return simplified item list or error. Args: query: Search query text. num_results: Number of results to request.

Searches Wikipedia based on the provided query and returns the results. Parameters: query (str): The search query for Wi

Executes the OCR tool to detect text in the provided image. Parameters: image (str): The path to the image file.

Generates and executes Python code based on the provided query. Parameters: query (str): A query describing the desired

Executes the arXiv search tool to find papers based on the given query. Parameters: query (str): The search query for ar

Executes the given Python code snippet. Parameters: code (str): The Python code snippet to be executed. Returns

Execute a tool with given parameters

Parse and execute tool from action. Returns (env_action, tool_result).

Extract final answers from agent responses. Looks for <answer>...</answer> tags or FINAL_ANSWER: format.

Extracts all text from the given URL. Parameters: url (str): The URL from which to extract text. Returns:

Fetch history with optional summarization. Strategy: - 1 step: return original history (no summarization needed)

Fetches a single page of news articles from Nature's website. Parameters: page_number (int): The page number to fetch.

Filter data by difficulty level if specified

Dynamic Sampling: Over-sample and filter out episode group in which all episodes have the same rewards. Adopted from DAPO (https://arxiv.

Generates a redeem code corresponding to the session ID for an MTurk worker once the session is completed

Generate mock staged responses.

Get action from LLM for a single observation

Get actions for multiple observations in parallel

Determines whether purchased products shares same attributes as goal

Returns list of available actions at the current step

random_colour_masks parameters: - image - predicted masks method: - the masks of each predicted o